Vigil — Development skill for Claude Code

Vigil

Agent-SDK TUI runner for gated validation plans and arbitrary prompts. Every tool call, gate verdict, hook decision, token and dollar is visible live in an animated cockpit — and persisted to disk so a run can be replayed or audited without the TUI.

Built on `@anthropic-ai/claude-agent-sdk` (the Claude Code agent loop) + Ink.

Install

cd ~/Desktop/vigil
npm install && npm run build && npm link      # → `vigil` on PATH

Requires Node ≥ 20 and a Claude Code executable (`claude` on PATH, `CLAUDE_CODE_EXECPATH`, or `--claude-bin`).

Use

vigil run ~/Desktop/OmniRoute/_tasks/planning/omp-live-deploy            # every plan, in order, halting on the first FAIL
vigil run …/omp-live-deploy --only 01-01                                  # one plan
vigil run …/omp-live-deploy --from 02-01 --to 02-02                       # a range
vigil prompt "Audit ~/.omniroute/logs for errors in the last hour"        # any prompt, same cockpit
vigil prompt -f brief.md --approvals all                                  # approve every tool call by hand
vigil replay /runs/ --speed 30                              # re-animate a recorded run
vigil tail   /runs/                                         # follow a run from another terminal
vigil runs                                                      # list runs with verdict + cost
vigil run … --plain                                                       # no TUI: readable line log (CI, ssh, logs)

Keys

`space` pause (holds the next tool call) · `i` inject an instruction into the live session · `x` interrupt the current turn · `a` cycle approvals auto → gates → all · `e` evidence overlay (criteria ‖ captured file) · `t` tool detail (full input/output, `j`/`k` to walk) · `l` diagnostics level · `f` filter by source · `r` re-run a failed plan · `q` quit.

What Vigil enforces (beyond what the agent says)

• Verdict lines. The runner contract appended to every p